Systematic review and comparison of machine learning and conventional statistical models for predicting cardiovascular events in dialysis patients

This systematic review aimed to evaluate the performance of machine learning (ML) models and conventional statistical models (CSMs) for predicting cardiovascular events in dialysis patients. Following PRISMA guidelines, eligible studies were searched through PubMed and Embase (January 2015–March 2025). Model performance (AUC/C-index) was compared using Mann–Whitney U test, while risk of bias was assessed via PROBAST. Furthermore, subgroup analyses stratified by algorithm type, validation method, and dataset size were conducted to explore heterogeneity. The review included 14 studies encompassing 29,310 patients and 34 models. Based on test/validation datasets only, ML models achieved comparable discrimination (mean AUC: 0.784 ± 0.112) than CSMs (0.772 ± 0.066), without statistical significance (p = 0.24). The PROBAST assessment indicated that 71.43% of studies had a low risk of bias. Subgroup analysis of performance revealed that deep learning models significantly outperformed both traditional ML and CSMs (p = 0.005), whereas traditional ML showed no advantage over CSMs (p = 0.727). Studies were predominantly originated from China (71.40%) and relied on internal validation (78.57%), limiting generalizability. Although deep learning algorithms show promises, ML models overall do not significantly outperform CSMs. CSMs remain viable, especially in resource-limited settings. Critical limitations include geographical bias, insufficient external validation, and tradeoffs between accuracy and interpretability. Future research should prioritize validation frameworks and clinical implementation over marginal accuracy improvements.

本系统综述旨在评估机器学习（Machine Learning, ML）模型与传统统计模型（Conventional Statistical Models, CSMs）在透析患者心血管事件预测任务中的表现。本研究遵循PRISMA指南，检索了2015年1月至2025年3月期间PubMed与Embase数据库中符合纳入标准的相关研究。模型性能（AUC/C-index）采用曼-惠特尼U检验（Mann–Whitney U test）进行比较，偏倚风险则通过PROBAST工具进行评估。此外，本研究还按照算法类型、验证方法及数据集规模开展亚组分析，以探究异质性来源。本综述共纳入14项研究，涵盖29310名患者与34个模型。仅基于测试/验证数据集的分析结果显示，机器学习模型的区分度（平均AUC：0.784±0.112）与传统统计模型（0.772±0.066）相当，且差异无统计学意义（p=0.24）。PROBAST评估结果表明，71.43%的纳入研究偏倚风险较低。性能亚组分析显示，深度学习模型的表现显著优于传统机器学习模型与传统统计模型（p=0.005），而传统机器学习模型并未展现出优于传统统计模型的优势（p=0.727）。纳入研究主要来自中国（71.40%），且多采用内部验证（78.57%），这限制了研究结果的外推性。尽管深度学习算法展现出应用前景，但整体而言机器学习模型并未显著优于传统统计模型。传统统计模型仍具备实用性，在资源受限场景中尤为如此。本综述的关键局限性包括地理偏倚、外部验证不足，以及准确性与可解释性之间的权衡。未来研究应优先关注验证框架搭建与临床落地应用，而非仅追求微小的精度提升。